Pith. sign in

REVIEW 4 major objections 6 minor 37 references

ThermoField recovers spatially varying thermal diffusivity on complex 3D objects from time-resolved surface temperatures and uses those fields to predict heat flow under new conditions.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A neural-field plus differentiable FEM pipeline recovers spatially varying thermal diffusivity on reconstructed 3D objects from synthetic thermal sequences and partially transfers to held-out heating/cooling conditions.

T0 review reviewed 2026-07-10 challenge →

load-bearing objection Solid methods bridge of neural surfaces + differentiable FEM for thermophysical fields on complex meshes; synthetic oracle evaluation and abstract overclaim are the real limits, not the core idea. the 4 major comments →

arxiv 2607.07962 v1 pith:UPGVBSLP submitted 2026-07-08 cs.CV cs.AI

Beyond Thermal Imaging: Inferring Thermophysical Properties from Time-Resolved Thermal Observations

Chenghao Xu , Malcolm Mielle , Olga Fink This is my paper
classification cs.CV cs.AI
keywords Thermal ImageryThermophysical PropertiesMaterial IdentificationDifferentiable Heat TransferNeural FieldsInverse Problems3D Scene Reconstruction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that thermal images should be treated as measurements of an underlying heat-transfer process, not as visual appearance to be reconstructed. It introduces ThermoField, which first reconstructs object geometry at metric scale, represents thermophysical quantities such as thermal diffusivity as neural fields on the surface, and then fits those fields by running a differentiable heat-equation solver until simulated surface temperatures match the observed sequence. On synthetic objects ranging from simple shapes to a bunny, bear, and car, the recovered fields are often close enough to ground truth that they can forecast temperature evolution under held-out heating or cooling without re-optimization. The claim matters because it turns thermal cameras into sensors of material properties rather than just temperature maps, supporting digital twins, infrastructure monitoring, and predictive simulation that stay consistent when the environment changes.

Core claim

ThermoField can jointly recover metrically scaled geometry and spatially varying thermophysical fields—primarily thermal diffusivity, and in some cases convective exchange coefficients—from time-resolved surface thermal observations of complex three-dimensional objects, and those fields remain predictive under previously unseen environmental conditions.

What carries the argument

ThermoField: neural fields of thermophysical quantities (diffusivity, scaled convection/radiation coefficients) defined on sparse surface control points and optimized by back-propagating temperature mismatch through a differentiable finite-element heat-transfer solver on the reconstructed tetrahedral mesh.

Load-bearing premise

The observed surface-temperature trajectories must carry enough independent information to pin down the target material field once geometry and the fixed boundary terms are given; when that information is weak, many different fields can still fit the same temperatures.

What would settle it

On a real or synthetic object whose true diffusivity is known, recover the field from one thermal process and then drive the forward simulator under a held-out heating or cooling condition; if predicted surface temperatures systematically diverge from measurement (or the recovered field stays far from ground truth despite low training error), the central claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. ThermoField proposes a physics-grounded inverse framework that recovers spatially varying thermophysical fields (primarily thermal diffusivity α, and in one case a scaled convection coefficient β) on metrically reconstructed 3D object surfaces from time-resolved surface temperature sequences. Geometry is first obtained via NeuS-style SDF reconstruction from multi-view RGB and converted to a tetrahedral mesh; neural fields on sparse surface control points then parameterize α(x)/β(x)/γ(x) and drive a differentiable JAX-FEM transient heat solver (Robin boundary conditions, staged TET4→TET10). Parameters are optimized by matching simulated to observed surface temperatures with smoothness regularization. On a synthetic ANSYS suite of six objects under cooling/heating/warming protocols, the method reports recovery accuracy (Table 1), cross-condition transfer (Table 2, Fig. 2), and multi-seed identifiability ensembles (Fig. 3), and discusses when passive cooling, high diffusivity, or symmetry leave parameters weakly constrained.

Significance. If the framework generalizes beyond the current synthetic regime, it would meaningfully bridge neural thermal scene representations and classical inverse heat transfer by enabling spatially resolved, predictive thermophysical fields on irregular reconstructed geometry rather than voxel grids or single global constants. Strengths include: an explicit differentiable FEM pipeline on reconstructed meshes; honest reporting of large recovery and transfer failures; multi-seed ensembles that separate observation-space fit from parameter uniqueness; and a clear staged discretization plus smoothness analysis. These are genuine contributions relative to thermal NeRF/GS methods that treat temperature as appearance and to inverse-HT methods restricted to simplified domains. The significance is currently limited by exclusive reliance on oracle synthetic data with known free/fixed coefficients and prescribed boundary schedules.

major comments (4)
  1. Abstract and opening claim state that ThermoField “jointly reconstructs geometry, estimates spatially varying thermal diffusivity, and predicts thermal evolution.” §4 and §4.1 instead separate the pipeline: geometry is reconstructed from multi-view RGB via NeuS, metrically scaled, then held fixed as the computational domain for a subsequent inverse solve. Thermal observations never enter geometry estimation. The abstract should be revised to match the actual two-stage design; “joint” reconstruction is not demonstrated.
  2. All quantitative support for the central claim (Tables 1–2, Figs. 2–3; Appendix A) is closed-loop ANSYS simulation with known material constants (Table A1), prescribed Ta/Q schedules (Table A2), and oracle choice of which coefficient is free (e.g., Cylinder–Cooling optimizes β while α is fixed to GT in the heating transfer test; §2.2). This tests whether differentiable FEM can re-identify simulator parameters under perfect boundary knowledge, not whether surface IR trajectories identify thermophysical fields under realistic radiometry, emissivity–temperature coupling, or unknown mixed fluxes—the non-identifiability sources §3 itself flags. At least one real thermal-camera experiment, or a controlled ablation with unknown free/fixed sets and boundary noise, is needed to substantiate the abstract’s claim for complex 3D scenes.
  3. §2.1–2.3 already show that the weakest assumption—sufficient independent information in surface T trajectories—often fails: Sphere–Cooling 42.8% relative error with a coherent but biased field; Bear–Heating 83.1% error and 307.8% normalized width; Cylinder–Cooling convection transfers to warming (MAE 0.038°C) but collapses under held-out heating (MAE ~45°C); ensembles yield low final-time MAE with non-unique fields (Fig. 3). The paper reports these results carefully, but the Abstract/Results framing still presents predictive transfer as demonstrated. Claims should be conditioned on excitation type, material class, and free-parameter choice, with explicit failure criteria rather than scene-by-scene narrative.
  4. §4.2–4.3 and Table 1: free vs fixed coefficients (α vs β vs γ), physical bounds [ϕ_min, ϕ_max], λ_smooth, control-point placement, and which process is used for training are chosen per scene with knowledge of the ground-truth material response. In a real inverse setting this oracle selection is unavailable. The manuscript should either (i) fix a single free-parameter protocol across all objects and report the resulting degradation, or (ii) provide an automatic model-selection / identifiability criterion before claiming general thermophysical inference.
minor comments (6)
  1. Typo in Abstract: “thermophyiscal” → “thermophysical”.
  2. Table 1 footnote: GT values “in units of 10^{-6} of the corresponding physical quantity” is easy to misread; state explicit units (e.g., α in 10^{-6} m²/s).
  3. Bunny is omitted from Table 2 because held-out conditions are undefined in Table A2; either add held-out Bunny configs or state this limitation once in the main text.
  4. §4.2 Eq. (3) introduces qb in the prose but not in the displayed equation; align notation with Appendix B Eq. (B19).
  5. Fig. 2 caption: clarify that “trained-sequence absolute error” is final-frame (cooling) vs end-of-heating-stage, so panels are not strictly comparable across process types.
  6. Appendix A.1: “observataiont”, “quaratic” typos; clean before camera-ready.

Circularity Check

0 steps flagged

No load-bearing circularity: inverse optimization of neural thermophysical fields via differentiable heat equation is standard and checked on held-out conditions and GT, not tautological.

full rationale

ThermoField parameterizes spatially varying α (or β) as neural fields on reconstructed surface control points, embeds them in a differentiable FEM heat-transfer solver (Eqs. 2–5, B18–B19), and optimizes by matching simulated surface temperatures to observed sequences (L_data, Eq. 6/B22) plus smoothness. This is ordinary physics-constrained inverse modeling, not a self-definitional loop: the recovered field is not algebraically identical to the observations or to any fitted scalar that is then re-labeled a prediction. Geometry is reconstructed separately (NeuS SDF) and held fixed; boundary terms are prescribed or selectively free; training uses one cooling + one heating sequence while testing uses held-out warming/heating with different T_a or Q̇ (Tables A1–A2, §2.2). Cross-condition MAE/RMSE and multi-seed ensembles (§2.3) are independent checks, and the paper itself reports non-uniqueness (low observation MAE coexisting with parameter scatter) rather than claiming uniqueness by construction. Self-citations (e.g., ThermoNeRF) are background, not load-bearing uniqueness theorems. Synthetic closed-loop evaluation with oracle boundaries is a validity limitation, not circularity of the derivation chain. Score 1 only for the minor risk that a reader might misread training-sequence fit as the claimed predictive result; the paper’s own held-out protocol prevents that reduction.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 1 invented entities

The claim rests on classical continuum heat transfer plus standard neural implicit geometry, plus several engineering choices (fixed vs free boundary coefficients, smoothness weight, physical box constraints, linearized radiation, metric scale from depth). Free parameters are regularization and representation knobs, not fitted universal constants. No new physical particle or force is postulated; “neural thermophysical fields” are a parameterization.

free parameters (6)
  • λ_smooth (field smoothness weight)
    Controls spatial coherence of recovered α/β; authors state removing it broadens 5–95% widths (§4.3). Chosen for optimization behavior, not measured.
  • λ_reg (neural parameter L2 weight)
    Standard weight decay on field network parameters in Eq. (8); hand-set.
  • Physical bounds [ϕ_min, ϕ_max] on α, β, γ
    Sigmoid-rescaled outputs constrained to prescribed ranges (§4.2); bounds shape the recoverable solution set.
  • Number/placement of surface control points and PE frequencies L
    Discretization of the neural field capacity; affects spatial resolution and overfitting (§B.5).
  • Which coefficients are free vs fixed per scene
    e.g., Cylinder–Cooling optimizes β with diffusivity fixed in transfer tests (Table 2 note); modeling choice that changes the inverse target.
  • Staged FEM schedule (TET4 then TET10) and Adam learning-rate scales
    Optimization hyperparameters that affect convergence basins (§4.3, §B.4, §B.8).
axioms (5)
  • domain assumption Transient temperature obeys the continuum heat equation with Fourier conduction and Robin convection/radiation boundary conditions on the reconstructed surface.
    Eqs. (2)–(3); standard continuum heat transfer assumed exact for the synthetic generator and inverse model.
  • ad hoc to paper Radiative flux may be linearized about the previous time step without destroying gradient-based recovery of the target fields.
    §4.3 / §B.3 numerical convenience; not independently validated against full nonlinear radiation for inverse bias.
  • domain assumption Metric-scale SDF geometry from multi-view RGB (NeuS + depth scale) is accurate enough that spatial derivatives in the heat equation remain faithful.
    §4.1; geometry is fixed before inversion; errors are absorbed into recovered fields (Discussion).
  • ad hoc to paper Surface thermal observations plus known external Ta/Q suffice to identify the chosen free field when complementary sequences are used.
    Central identifiability hypothesis of §2; contradicted in several reported scenes, so it is an assumption of the claim’s scope, not a theorem.
  • standard math Standard automatic differentiation through implicit FEM time stepping yields correct gradients for inverse optimization.
    JAX-FEM pipeline §4.3; relies on correctness of the differentiable solver implementation.
invented entities (1)
  • ThermoField neural thermophysical surface fields (α(x), β(x), γ(x) via MLP on control points) no independent evidence
    purpose: Parameterize spatially varying material/boundary coefficients on irregular reconstructed meshes for gradient-based inverse heat transfer.
    Representation device, not a new physical substance; independent evidence would be real-material measurements matching recovered maps—only synthetic GT is shown.

reviewed 2026-07-10 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Thermal Imaging: Inferring Thermophysical Properties from Time-Resolved Thermal Observations." pith.science (2026). https://pith.science/paper/UPGVBSLP

@misc{pith2026260707962,
  author       = {Pith},
  title        = {Pith review of: Beyond Thermal Imaging: Inferring Thermophysical Properties from Time-Resolved Thermal Observations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UPGVBSLP}},
  note         = {Machine review of arXiv:2607.07962}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Inferring latent physical properties from sensory observations is a fundamental challenge in machine perception. Among available sensing modalities, thermal imaging is particularly promising because temperature evolution is directly governed by heat-transfer physics and therefore encodes information about underlying thermophysical properties of a scene. Recovering spatially resolved thermophysical properties from thermal observations could transform applications ranging from digital twins and infrastructure monitoring to robotics and scientific imaging. However, existing thermal scene reconstruction methods can recover temperature fields in complex 3D environments without identifying the thermophyiscal properties that govern thermal evolution, whereas inverse methods provide physically interpretable parameter estimation but typically rely on simplified geometries and controlled experimental conditions. Here we introduce ThermoField, a framework that unifies thermal scene reconstruction and thermophysical parameter estimation through differentiable heat-transfer simulation. The proposed framework represents these quantities as spatially varying neural fields and constrains them through scene geometry, governing heat-transfer physics, and temporal thermal observations. We demonstrate that ThermoField jointly reconstructs geometry, estimates spatially varying thermal diffusivity, and predicts thermal evolution under previously unseen environmental conditions. By integrating neural scene representations with differentiable heat-transfer solver, the framework enables physically interpretable parameter inference in complex 3D scenes. Our results establish a bridge between thermal scene reconstruction and inverse heat-transfer analysis, providing a unified approach for geometry reconstruction, thermophysical property estimation, and predictive thermal simulation from thermal observations.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 37 canonical work pages · 2 internal anchors

  1. [1]

    (2024) 19

    Zheng, Y., Zhao, Q., Yang, G., Yifan, W., Xiang, D., Dubost, F., Lagun, D., Beeler, T., Tombari, F., Guibas, L., Wetzstein, G.: Physavatar: Learning the physics of dressed 3d avatars from visual observations. (2024) 19

    work page 2024

  2. [2]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Li, W., Gu, Y., Chen, X., Xu, X., Hu, M., Huang, X., Wu, Y.: Towards visual discrimination and reasoning of real-world physical dynamics: Physics-grounded anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 30409–30419 (2025)

    work page 2025

  3. [3]

    In: Proceedings of the 37th International Conference on Neural Information Processing Systems

    Tung, H.-Y., Ding, M., Chen, Z., Bear, D.M., Gan, C., Tenenbaum, J.B., Yamins, D.L.K., Fan, J., Smith, K.A.: Physion++: evaluating physical scene understand- ing that requires online inference of different physical properties. In: Proceedings of the 37th International Conference on Neural Information Processing Systems. NIPS ’23. Curran Associates Inc., R...

    work page 2023

  4. [4]

    In: The Thirteenth International Conference on Learning Representations (ICLR) (2025)

    Chow, W., Mao, J., Li, B., Seita, D., Guizilini, V., Wang, Y.: Physbench: Bench- marking and enhancing vision-language models for physical world understanding. In: The Thirteenth International Conference on Learning Representations (ICLR) (2025)

    work page 2025

  5. [5]

    https://arxiv.org/abs/2601.01321

    Zhou, R., Chen, D., Jia, Z., Su, Y., Liu, Y., Lu, Y., Shi, D., Huang, Y., Xu, T., Pan, Y., Li, X., Abate, Y., Chen, Q., Tu, Z., Yang, Y., Zhang, Y., Wen, Q., Mai, G., Fu, S., Li, J., Wang, X., Wang, Z., Huang, J., Liu, T., Chen, Y., Sun, L., He, L.: Digital Twin AI: Opportunities and Challenges from Large Language Models to World Models (2026). https://ar...

    work page arXiv 2026

  6. [6]

    In: Gon¸ calves, N., Oliveira, H.P., S´ anchez, J.A

    Cardoso, R.P., Moreno, P.: Estimating object physical properties from rgb-d vision and depth robot sensors using deep learning. In: Gon¸ calves, N., Oliveira, H.P., S´ anchez, J.A. (eds.) Pattern Recognition and Image Analysis, pp. 97–110. Springer, Cham (2026)

    work page 2026

  7. [7]

    IEEE Robotics and Automation Letters9(11), 9279–9286 (2024) https://doi.org/10.1109/LRA.2024.3455788

    Chen, J., Frey, J., Zhou, R., Miki, T., Martius, G., Hutter, M.: Identifying terrain physical parameters from vision - towards physical-parameter-aware locomotion and navigation. IEEE Robotics and Automation Letters9(11), 9279–9286 (2024) https://doi.org/10.1109/LRA.2024.3455788

    work page doi:10.1109/lra.2024.3455788 2024

  8. [8]

    In: IEEE International Conference on Robotics and Automation (ICRA) (2024)

    Gao, J., Sarkar, B., Xia, F., Xiao, T., Wu, J., Ichter, B., Majumdar, A., Sadigh, D.: Physically grounded vision-language models for robotic manipulation. In: IEEE International Conference on Robotics and Automation (ICRA) (2024). IEEE

    work page 2024

  9. [9]

    Efficient Physics Simulation for 3D Scenes via MLLM-Guided Gaussian Splatting

    Zhao, H., Wang, H., Zhao, X., Fei, H., Wang, H., Long, C., Zou, H.: Effi- cient physics simulation for 3d scenes via mllm-guided gaussian splatting. arXiv preprint arXiv:2411.12789 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  10. [10]

    Jr: An inverse problem in thermal imaging

    Bryan, K., Caudill, L.F. Jr: An inverse problem in thermal imaging. SIAM journal on Applied Mathematics56(3), 715–735 (1996)

    work page 1996

  11. [11]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Dashpute, A., Saragadam, V., Alexander, E., Willomitzer, F., Katsaggelos, A., Veeraraghavan, A., Cossairt, O.: Thermal spread functions (tsf): Physics- guided material classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 1641–1650 (2023) 20

    work page 2023

  12. [12]

    Computer Graphics Forum22(4), 663–687 (2003) https://doi.org/10.1111/j.1467-8659.2003.00716.x https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1467-8659.2003.00716.x

    Patow, G., Pueyo, X.: A survey of inverse rendering problems. Computer Graphics Forum22(4), 663–687 (2003) https://doi.org/10.1111/j.1467-8659.2003.00716.x https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1467-8659.2003.00716.x

    work page doi:10.1111/j.1467-8659.2003.00716.x 2003

  13. [13]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Li, C., Ono, T., Uemori, T., Mihara, H., Gatto, A., Nagahara, H., Moriuchi, Y.: Neisf: Neural incident stokes field for geometry and material estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 21434–21445 (2024)

    work page 2024

  14. [14]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025)

    Wu, S., Basu, S., Broedermann, T., Van Gool, L., Sakaridis, C.: PBR-NeRF: Inverse rendering with physics-based neural fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025)

    work page 2025

  15. [15]

    In: Computer Vision and Pattern Recognition (CVPR) (2019)

    Kanazawa, A., Zhang, J.Y., Felsen, P., Malik, J.: Learning 3d human dynamics from video. In: Computer Vision and Pattern Recognition (CVPR) (2019)

    work page 2019

  16. [16]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference, pp

    Chen, C., Dou, Z., Wang, C., Huang, Y., Chen, A., Feng, Q., Gu, J., Liu, L.: Vid2sim: Generalizable, video-based reconstruction of appearance, geometry and physics for mesh-free simulation. In: Proceedings of the Computer Vision and Pattern Recognition Conference, pp. 26545–26555 (2025)

    work page 2025

  17. [17]

    In: Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVII, pp

    Purri, M., Dana, K.: Teaching cameras to feel: Estimating tactile physical properties of surfaces from images. In: Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVII, pp. 1–20. Springer, Berlin, Heidelberg (2020). https://doi.org/10.1007/ 978-3-030-58583-9 1

    work page 2020

  18. [18]

    Physics of Thermal Radiation, pp

    DeWitt, D.P., Incropera, F.P.: 1. Physics of Thermal Radiation, pp. 19–89. John Wiley I& Sons, Ltd, ??? (1988). https://doi.org/10.1002/9780470172575.ch1 . https://onlinelibrary.wiley.com/doi/abs/10.1002/9780470172575.ch1

    work page doi:10.1002/9780470172575.ch1 1988

  19. [19]

    Some Basic Concepts in Heat Transfer, pp. 351–392. John Wiley I& Sons, Ltd (2017). Chap. 4. https://doi.org/10.1002/9783527693306.ch4 . https:// onlinelibrary.wiley.com/doi/abs/10.1002/9783527693306.ch4

    work page doi:10.1002/9783527693306.ch4 2017

  20. [20]

    In: Ferrarini, G., Spaeth, P., L´ opez, F

    Krapez, J.-C.: Multiwavelength thermometry without a priori emissivity infor- mation: from promise to disillusionment. In: Ferrarini, G., Spaeth, P., L´ opez, F. (eds.) Thermosense: Thermal Infrared Applications XLVII, vol. 13470, p. 1347006. SPIE, ??? (2025). https://doi.org/10.1117/12.3065037 . International Society for Optics and Photonics.https://doi....

    work page doi:10.1117/12.3065037 2025

  21. [21]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Narayanan, S., Ramanagopal, M., Narasimhan, S.: Dual band thermal videog- raphy: Separating time-varying reflection and emission near ambient conditions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 199–208 (2026) 21

    work page 2026

  22. [22]

    Engi- neering Applications of Artificial Intelligence118, 105640 (2023) https://doi.org/ 10.1016/j.engappai.2022.105640

    Wang, H., Song, K., Huang, L., Wen, H., Yan, Y.: Thermal images-aware guided early fusion network for cross-illumination rgb-t salient object detection. Engi- neering Applications of Artificial Intelligence118, 105640 (2023) https://doi.org/ 10.1016/j.engappai.2022.105640

    work page doi:10.1016/j.engappai.2022.105640 2023

  23. [23]

    IEEE Transactions on Automation Science and Engineering18(3), 1000–1011 (2021) https://doi.org/ 10.1109/TASE.2020.2993143

    Sun, Y., Zuo, W., Yun, P., Wang, H., Liu, M.: Fuseseg: Semantic segmentation of urban scenes based on rgb and thermal data fusion. IEEE Transactions on Automation Science and Engineering18(3), 1000–1011 (2021) https://doi.org/ 10.1109/TASE.2020.2993143

    work page doi:10.1109/tase.2020.2993143 2021

  24. [24]

    Automation in Construction168, 105709 (2024)

    Vollmer, E., Ruck, J., Volk, R., Schultmann, F.: Detecting district heating leaks in thermal imagery: Comparison of anomaly detection methods. Automation in Construction168, 105709 (2024)

    work page 2024

  25. [25]

    Advanced Engineering Informatics65, 103345 (2025)

    Hassan, M., Forest, F., Fink, O., Mielle, M.: Thermonerf: A multimodal neu- ral radiance field for joint rgb-thermal novel view synthesis of building facades. Advanced Engineering Informatics65, 103345 (2025)

    work page 2025

  26. [26]

    In: European Conference on Computer Vision (2024)

    Chen, Q., Shu, s., Bai, X.: Thermal3d-gs: Physics-induced 3d gaussians for ther- mal infrared novel-view synthesis. In: European Conference on Computer Vision (2024)

    work page 2024

  27. [27]

    Frontiers in neurorobotics13, 6 (2019)

    Degrave, J., Hermans, M., Dambre, J., Wyffels, F.: A differentiable physics engine for deep learning in robotics. Frontiers in neurorobotics13, 6 (2019)

    work page 2019

  28. [28]

    https://arxiv.org/abs/2511.23449

    Waseem, A., Mielle, M.: Physics-Informed Neural Networks for Thermophysical Property Retrieval (2025). https://arxiv.org/abs/2511.23449

    work page arXiv 2025

  29. [29]

    ICLR (2020)

    Hu, Y., Anderson, L., Li, T.-M., Sun, Q., Carr, N., Ragan-Kelley, J., Durand, F.: Difftaichi: Differentiable programming for physical simulation. ICLR (2020)

    work page 2020

  30. [30]

    In: Cortes, C., Lawrence, N., Lee, D., Sugiyama, M., Garnett, R

    Wu, J., Yildirim, I., Lim, J.J., Freeman, B., Tenenbaum, J.: Galileo: Perceiving physical object properties by integrating a physics engine with deep learning. In: Cortes, C., Lawrence, N., Lee, D., Sugiyama, M., Garnett, R. (eds.) Advances in Neural Information Processing Systems, vol. 28. Curran Associates, Inc., ??? (2015)

    work page 2015

  31. [31]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pp

    Yang, K., Liu, Y., Cui, Z., Liu, Y., Zhang, M., Yan, S., Wang, Q.: Ntr- gaussian: Nighttime dynamic thermal reconstruction with 4d gaussian splatting based on thermodynamics. In: Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pp. 691–700 (2025)

    work page 2025

  32. [32]

    In: The Fourteenth International Conference on Learning Representations 22

    Wang, Z., Ling, H., Zhang, W., Sun, Y., Sun, Q.: Etgs: Explicit thermodynam- ics gaussian splatting for dynamic thermal reconstruction. In: The Fourteenth International Conference on Learning Representations 22

    work page

  33. [33]

    In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Saponaro, P., Sorensen, S., Kolagunda, A., Kambhamettu, C.: Material classifi- cation with thermal imagery. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4649–4656 (2015). https://doi.org/10.1109/ CVPR.2015.7299096

    work page arXiv 2015

  34. [34]

    ANSYS, Inc., (2024)

    ANSYS, Inc.: Ansys Mechanical User’s Guide. ANSYS, Inc., (2024). ANSYS, Inc.. Ansys Mechanical, Release 2024 R2, Help System. https://ansyshelp.ansys. com/public/Views/Secured/corp/v242/en/wb sim/ds Home.html

    work page 2024

  35. [35]

    NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view Reconstruction

    Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  36. [36]

    In: Proceedings of Machine Learning and Systems 2020, pp

    Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric reg- ularization for learning shapes. In: Proceedings of Machine Learning and Systems 2020, pp. 3569–3579 (2020)

    work page 2020

  37. [37]

    In: Proceedings of the 14th Annual Conference on Computer Graphics and Interactive Techniques

    Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface con- struction algorithm. In: Proceedings of the 14th Annual Conference on Computer Graphics and Interactive Techniques. SIGGRAPH ’87, pp. 163–169. Association for Computing Machinery, New York, NY, USA (1987). https://doi.org/10.1145/ 37401.37422 .https://doi.org/10.1145/37401.3742...

    work page doi:10.1145/37401.37422 1987

This paper was first reviewed by grok-4.5 on July 10, 2026.